2025-07-29

AI Expert Panel: America’s AI Plan, the End of Google Search & the Next ChatGPT | EP #185

Peter Diamandis hosts Dave Blundin and first-time co-host Alex Wissner-Gross (Salim Ismail is away) for a rapid-fire tour of the AI wars: the GPU buildout race between xAI, OpenAI, and Meta, Trump's newly unveiled America's AI Action Plan, and whether frontier benchmarks (IMO gold, Humanity's Last Exam) are already saturating. Wissner-Gross argues AGI has effectively existed since GPT-3 in 2020 and that math and physics are close to being 'solved' by AI, while the group debates US energy bottlenecks versus China's solar buildout, agentic browsers replacing search, and the societal fallout of AI companions among teens. The episode closes on quantum computing, robotaxis as mobile data centers, and Tesla's new AI-themed diner as a symbol of science fiction becoming real.

▶ Watch on YouTube

Topics

GPU/data center buildout race Compute ▶ 0:10

xAI's Colossus, OpenAI's GPU targets, and Meta's Prometheus/Hyperion data centers are compared; panel discusses GB200 chip economics, hurricane-proof tent data centers, and novel form factors like ocean- or vehicle-based compute.

America's AI Action Plan Geopolitics ▶ 56:34

Trump's 90+ policy AI action plan deregulates AI, fast-tracks data center/chip/energy permitting, and funds nuclear and geothermal power; compared to the Manhattan Project and the interstate highway system.

US vs China energy strategy Energy ▶ 1:41:42

China installed 464 GW of solar in 12 months while the US leans on nuclear/geothermal/natural gas; panel debates why the US isn't scaling solar faster despite abundant capacity sitting unused.

Benchmark saturation and IMO gold AI ▶ 37:12

OpenAI and Google DeepMind both won gold at the 2025 International Math Olympiad using natural-language reasoning (no formalization step); Wissner-Gross argues Humanity's Last Exam, AIME, GPQA and SWE-bench are all saturating and the field needs new 'open problem' benchmarks.

AGI already here debate AI ▶ 43:32

Wissner-Gross argues AGI has existed since GPT-3's few-shot learning paper (May 2020) as a smooth exponential rather than a discrete moment, a claim Blundin endorses.

AI talent wars and compensation Economy ▶ 18:23

Meta's superintelligence team (half Chinese-origin, poached heavily from OpenAI/DeepMind/Scale) is paid $10M-$100M+/year; the panel discusses whether individual human comp at this level is sustainable as AI automates labor.

Browser wars and agentic search AI ▶ 1:18:32

OpenAI's upcoming Chromium-based browser and Perplexity's Comet challenge Google Chrome; panel argues the 'browser' concept itself is becoming obsolete in favor of agentic portals and computer-use agents.

Google search revenue vs AI overviews AI ▶ 1:10:36

A WSJ report claims AI overviews are helping (not hurting) Google search revenue ($54.2B in Q2); Blundin disputes this, arguing AI overviews are cannibalizing clickable links and search is moving entirely to AI.

Recursive self-improvement and AI safety AI ▶ 1:25:00

Sparked by AI writing ~50% of Google's code, the panel debates recursive self-improvement risk; Blundin argues simple guardrails (don't let AI set its own objectives) neutralize the danger, while Wissner-Gross leans accelerationist and worries more about overregulation.

AI companions and teen mental health Health ▶ 1:35:51

73% of teens 14-17 have used an AI companion, with usage correlated to depression and loneliness; Blundin argues the addictive design incentives mirror video games since there's no business incentive to encourage breaks.

Quantum computing and AI Compute ▶ 49:32

Panel discusses a scalable quantum computer arriving within ~3 years, whether neural networks can be trained on quantum hardware, and Blundin/Wissner-Gross's plan to assemble a team to answer this.

Robotaxis as mobile data centers Transportation ▶ 1:48:15

Uber invests $300M in Lucid, buying 20,000 Gravity EVs to compete with Tesla and Waymo; Wissner-Gross reframes autonomous vehicles as 'mobile distributed data centers' rather than smartphones on wheels.

Predictions made

open Elon Musk: xAI will scale to 50 million H100-equivalent GPUs.

EP #? · · due: 2030 · ▶ watch

“his goal is 50 million H100 equivalents within the next 5 years”

Your call:

open Sam Altman: OpenAI will have over 1 million GPUs brought online.

EP #? · · due: end of 2025 · ▶ watch

“We will cross well over 1 million GPUs brought online by the end of this year.”

Your call:

open Alex Wissner-Gross: 20% of the hardest mathematician-level problems will be solvable by AI, rising to 60-70% within two years, at which point math is effectively 'solved.'

EP #? · · due: 2027 · ▶ watch

“we're going to see 20% of mathematician level hard problems be solvable by the end of this year... and potentially 60% or 70% of them be solvable in the next 2 years”

Your call:

open Dave Blundin: A scalable, functional quantum computer will exist.

EP #? · · due: 2028 (within 3 years) · ▶ watch

“a scalable functional quantum computer is within 3 years according to our our best experts around MIT, maybe even sooner”

Your call:

open Dave Blundin: Data center buildout capex will reach roughly $1 trillion per year.

EP #? · · due: 2029 · ▶ watch

“the data center buildout is going to be a trillion dollars a year starting in 2029”

Your call:

open Dave Blundin: How people buy goods (and make life decisions like travel) will be fundamentally transformed by AI agents.

EP #? · · due: 12-24 months from July 2025 · ▶ watch

“how we buy stuff is going to be fundamentally transformed over the next 12 to 24 months. Fundamental.”

Your call:

open Alex Wissner-Gross: 'Abundance benchmarks' tied to human welfare outcomes will emerge and, once mechanically verifiable, get solved by AI at a high rate.

EP #? · · due: 2027-2028 (2-3 years) · ▶ watch

“I I would predict within 2 to 3 years of the benchmarks if they can be uh mechanically verified by uh benchmarking organizations you'll see a lot of those problems get solved by AI”

Your call:

open Alex Wissner-Gross: Several new state-of-the-art frontier models will emerge, more labs will achieve IMO Gold, and at least one grand-challenge-level math or science problem will be solved by AI.

EP #? · · due: next couple of months (late 2025) · ▶ watch

“I want to see several new state-of-the-art frontier models... I'd like to see ideally at least one grand challenge level problem in math or some other physical science get solved by AI.”

Your call:

Numbers that matter

xAI Colossus 2 has the equivalent of 5.5 million H100s (550 GB200 units) GB200 has two Blackwell GPUs each; production numbers must be divided by two to compare to H100 counts.
GB200 chips cost $60,000 each Cited as a reason not to deploy them on mobile/vehicle platforms carelessly.
Meta superintelligence team compensation ranges from $10 million to $100+ million per year 75% of the team holds PhDs; team sourced 50% from China, 40% ex-OpenAI, 20% ex-DeepMind, 15% from the Scale AI acquisition.
Anthropic revenue surged from $3 billion to $4 billion run-rate in one month; Claude generates $200 million with 60% margins from coding Cited alongside reports of Anthropic being valued around $100+ billion amid Amazon partnership and Apple acquisition rumors.
Frontier Math Tier 4 benchmark performance trend suggests 20% of hardest problems solvable by end of 2025, 60-70% within 2 years Based on extrapolating recent rate of improvement on the hardest professional-mathematician-level problems.
OpenAI and Google DeepMind both scored 35/42 at IMO 2025, solving 5 of 6 problems The US team scored 33-39/42 (second place); the Chinese team scored a perfect 42/42 across all six members.
Google Q2 2025 search revenue hit $54.2 billion, with AI overviews reaching 2+ billion users/month and search impressions up 49% Cited from a WSJ article arguing AI overviews help rather than hurt Google search; disputed by Dave Blundin as masking cannibalization.
AI now writes 50% of Google's code, up from 25% two years ago; Amazon 25%, Microsoft 25-30%, Robinhood 50% Used to debate proximity to recursive self-improvement.
73% of teens aged 14-17 have used an AI companion; 37% shared personal secrets; users are 2x more likely to feel depressed or lonely Raised as a mental-health/socialization concern.
Nigerian GPT-4 pilot delivered two years of learning in two weeks, a 1,200% acceleration Contrasted with slow AI adoption in US education.
China installed 464 gigawatts of solar capacity in the 12 months since June 2024 Contrasted with the US's slower nuclear/geothermal/gas-focused energy strategy.
A Chinese fusion reactor sustained superheated plasma for 1,066 seconds at 180 million°F Cited as a fusion milestone, though Wissner-Gross notes fusion converts less than 1% of reactant rest mass to energy.
Uber invested $300 million in Lucid, agreeing to buy 20,000 Gravity EVs over six years Part of Uber's robotaxi strategy to compete with Tesla and Waymo.
Trump announced $92 billion in AI/energy investments from 20 companies in Pennsylvania Shown via video clip promoting the America's AI Action Plan rollout.

Worth digging into

🕳️ Abundance benchmarks

Wissner-Gross and Diamandis propose replacing saturating benchmarks (HLE, AIME) with 'abundance benchmarks' tied to curing disease or solving energy, a framing that could redirect AI research incentives industry-wide.

🕳️ Neural networks on quantum computers

Blundin says he and Wissner-Gross are assembling a team specifically to determine whether neural network training is a domain where quantum computers provide real speedup, calling it a 'very profitable pathway.'

🕳️ AGI-since-2020 thesis

Wissner-Gross's claim that AGI has existed since GPT-3's May 2020 few-shot learning paper reframes the entire AGI timeline debate as a smooth exponential rather than a discrete event.

🕳️ Dyson swarm absence as evidence against solar-scaling superintelligence

Wissner-Gross uses the Fermi-paradox-style absence of detected Dyson swarms in infrared surveys to argue solar-driven compute growth won't scale indefinitely toward superintelligence, linking astrobiology search results to near-term energy strategy.

🕳️ Grounding cost economics as an unsolved bottleneck

Alex flags that grounding (search-verifying LLM answers) costs haven't fallen the way raw model capability costs have, and calls it the next frontier for a disruptive non-incumbent company.